LoopUnswitch.cpp revision 228ebd0f4cf9207d32d61ef4b11b81736895dc09
1//===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass transforms loops that contain branches on loop-invariant conditions 11// to have multiple loops. For example, it turns the left into the right code: 12// 13// for (...) if (lic) 14// A for (...) 15// if (lic) A; B; C 16// B else 17// C for (...) 18// A; C 19// 20// This can increase the size of the code exponentially (doubling it every time 21// a loop is unswitched) so we only unswitch if the resultant code will be 22// smaller than a threshold. 23// 24// This pass expects LICM to be run before it to hoist invariant conditions out 25// of the loop, to make the unswitching opportunity obvious. 26// 27//===----------------------------------------------------------------------===// 28 29#define DEBUG_TYPE "loop-unswitch" 30#include "llvm/Transforms/Scalar.h" 31#include "llvm/Constants.h" 32#include "llvm/DerivedTypes.h" 33#include "llvm/Function.h" 34#include "llvm/Instructions.h" 35#include "llvm/LLVMContext.h" 36#include "llvm/Analysis/ConstantFolding.h" 37#include "llvm/Analysis/InlineCost.h" 38#include "llvm/Analysis/LoopInfo.h" 39#include "llvm/Analysis/LoopPass.h" 40#include "llvm/Analysis/Dominators.h" 41#include "llvm/Transforms/Utils/Cloning.h" 42#include "llvm/Transforms/Utils/Local.h" 43#include "llvm/Transforms/Utils/BasicBlockUtils.h" 44#include "llvm/ADT/Statistic.h" 45#include "llvm/ADT/SmallPtrSet.h" 46#include "llvm/ADT/STLExtras.h" 47#include "llvm/Support/CommandLine.h" 48#include "llvm/Support/Debug.h" 49#include "llvm/Support/raw_ostream.h" 50#include <algorithm> 51#include <set> 52using namespace llvm; 53 54STATISTIC(NumBranches, "Number of branches unswitched"); 55STATISTIC(NumSwitches, "Number of switches unswitched"); 56STATISTIC(NumSelects , "Number of selects unswitched"); 57STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 58STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 59 60// The specific value of 50 here was chosen based only on intuition and a 61// few specific examples. 62static cl::opt<unsigned> 63Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 64 cl::init(50), cl::Hidden); 65 66namespace { 67 class LoopUnswitch : public LoopPass { 68 LoopInfo *LI; // Loop information 69 LPPassManager *LPM; 70 71 // LoopProcessWorklist - Used to check if second loop needs processing 72 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 73 std::vector<Loop*> LoopProcessWorklist; 74 SmallPtrSet<Value *,8> UnswitchedVals; 75 76 bool OptimizeForSize; 77 bool redoLoop; 78 79 Loop *currentLoop; 80 DominanceFrontier *DF; 81 DominatorTree *DT; 82 BasicBlock *loopHeader; 83 BasicBlock *loopPreheader; 84 85 // LoopBlocks contains all of the basic blocks of the loop, including the 86 // preheader of the loop, the body of the loop, and the exit blocks of the 87 // loop, in that order. 88 std::vector<BasicBlock*> LoopBlocks; 89 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 90 std::vector<BasicBlock*> NewBlocks; 91 92 public: 93 static char ID; // Pass ID, replacement for typeid 94 explicit LoopUnswitch(bool Os = false) : 95 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false), 96 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL), 97 loopPreheader(NULL) {} 98 99 bool runOnLoop(Loop *L, LPPassManager &LPM); 100 bool processCurrentLoop(); 101 102 /// This transformation requires natural loop information & requires that 103 /// loop preheaders be inserted into the CFG... 104 /// 105 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 106 AU.addRequiredID(LoopSimplifyID); 107 AU.addPreservedID(LoopSimplifyID); 108 AU.addRequired<LoopInfo>(); 109 AU.addPreserved<LoopInfo>(); 110 AU.addRequiredID(LCSSAID); 111 AU.addPreservedID(LCSSAID); 112 AU.addPreserved<DominatorTree>(); 113 AU.addPreserved<DominanceFrontier>(); 114 } 115 116 private: 117 118 virtual void releaseMemory() { 119 UnswitchedVals.clear(); 120 } 121 122 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 123 /// remove it. 124 void RemoveLoopFromWorklist(Loop *L) { 125 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 126 LoopProcessWorklist.end(), L); 127 if (I != LoopProcessWorklist.end()) 128 LoopProcessWorklist.erase(I); 129 } 130 131 void initLoopData() { 132 loopHeader = currentLoop->getHeader(); 133 loopPreheader = currentLoop->getLoopPreheader(); 134 } 135 136 /// Split all of the edges from inside the loop to their exit blocks. 137 /// Update the appropriate Phi nodes as we do so. 138 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 139 140 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 141 unsigned getLoopUnswitchCost(Value *LIC); 142 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 143 BasicBlock *ExitBlock); 144 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 145 146 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 147 Constant *Val, bool isEqual); 148 149 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 150 BasicBlock *TrueDest, 151 BasicBlock *FalseDest, 152 Instruction *InsertPt); 153 154 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 155 void RemoveBlockIfDead(BasicBlock *BB, 156 std::vector<Instruction*> &Worklist, Loop *l); 157 void RemoveLoopFromHierarchy(Loop *L); 158 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 159 BasicBlock **LoopExit = 0); 160 161 }; 162} 163char LoopUnswitch::ID = 0; 164static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 165 166Pass *llvm::createLoopUnswitchPass(bool Os) { 167 return new LoopUnswitch(Os); 168} 169 170/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 171/// invariant in the loop, or has an invariant piece, return the invariant. 172/// Otherwise, return null. 173static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 174 // Constants should be folded, not unswitched on! 175 if (isa<Constant>(Cond)) return 0; 176 177 // TODO: Handle: br (VARIANT|INVARIANT). 178 179 // Hoist simple values out. 180 if (L->makeLoopInvariant(Cond, Changed)) 181 return Cond; 182 183 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 184 if (BO->getOpcode() == Instruction::And || 185 BO->getOpcode() == Instruction::Or) { 186 // If either the left or right side is invariant, we can unswitch on this, 187 // which will cause the branch to go away in one loop and the condition to 188 // simplify in the other one. 189 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 190 return LHS; 191 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 192 return RHS; 193 } 194 195 return 0; 196} 197 198bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 199 LI = &getAnalysis<LoopInfo>(); 200 LPM = &LPM_Ref; 201 DF = getAnalysisIfAvailable<DominanceFrontier>(); 202 DT = getAnalysisIfAvailable<DominatorTree>(); 203 currentLoop = L; 204 Function *F = currentLoop->getHeader()->getParent(); 205 bool Changed = false; 206 do { 207 assert(currentLoop->isLCSSAForm()); 208 redoLoop = false; 209 Changed |= processCurrentLoop(); 210 } while(redoLoop); 211 212 if (Changed) { 213 // FIXME: Reconstruct dom info, because it is not preserved properly. 214 if (DT) 215 DT->runOnFunction(*F); 216 if (DF) 217 DF->runOnFunction(*F); 218 } 219 return Changed; 220} 221 222/// processCurrentLoop - Do actual work and unswitch loop if possible 223/// and profitable. 224bool LoopUnswitch::processCurrentLoop() { 225 bool Changed = false; 226 LLVMContext &Context = currentLoop->getHeader()->getContext(); 227 228 // Loop over all of the basic blocks in the loop. If we find an interior 229 // block that is branching on a loop-invariant condition, we can unswitch this 230 // loop. 231 for (Loop::block_iterator I = currentLoop->block_begin(), 232 E = currentLoop->block_end(); 233 I != E; ++I) { 234 TerminatorInst *TI = (*I)->getTerminator(); 235 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 236 // If this isn't branching on an invariant condition, we can't unswitch 237 // it. 238 if (BI->isConditional()) { 239 // See if this, or some part of it, is loop invariant. If so, we can 240 // unswitch on it if we desire. 241 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 242 currentLoop, Changed); 243 if (LoopCond && UnswitchIfProfitable(LoopCond, 244 ConstantInt::getTrue(Context))) { 245 ++NumBranches; 246 return true; 247 } 248 } 249 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 250 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 251 currentLoop, Changed); 252 if (LoopCond && SI->getNumCases() > 1) { 253 // Find a value to unswitch on: 254 // FIXME: this should chose the most expensive case! 255 Constant *UnswitchVal = SI->getCaseValue(1); 256 // Do not process same value again and again. 257 if (!UnswitchedVals.insert(UnswitchVal)) 258 continue; 259 260 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 261 ++NumSwitches; 262 return true; 263 } 264 } 265 } 266 267 // Scan the instructions to check for unswitchable values. 268 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 269 BBI != E; ++BBI) 270 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 271 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 272 currentLoop, Changed); 273 if (LoopCond && UnswitchIfProfitable(LoopCond, 274 ConstantInt::getTrue(Context))) { 275 ++NumSelects; 276 return true; 277 } 278 } 279 } 280 return Changed; 281} 282 283/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 284/// 1. Exit the loop with no side effects. 285/// 2. Branch to the latch block with no side-effects. 286/// 287/// If these conditions are true, we return true and set ExitBB to the block we 288/// exit through. 289/// 290static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 291 BasicBlock *&ExitBB, 292 std::set<BasicBlock*> &Visited) { 293 if (!Visited.insert(BB).second) { 294 // Already visited and Ok, end of recursion. 295 return true; 296 } else if (!L->contains(BB)) { 297 // Otherwise, this is a loop exit, this is fine so long as this is the 298 // first exit. 299 if (ExitBB != 0) return false; 300 ExitBB = BB; 301 return true; 302 } 303 304 // Otherwise, this is an unvisited intra-loop node. Check all successors. 305 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 306 // Check to see if the successor is a trivial loop exit. 307 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 308 return false; 309 } 310 311 // Okay, everything after this looks good, check to make sure that this block 312 // doesn't include any side effects. 313 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 314 if (I->mayHaveSideEffects()) 315 return false; 316 317 return true; 318} 319 320/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 321/// leads to an exit from the specified loop, and has no side-effects in the 322/// process. If so, return the block that is exited to, otherwise return null. 323static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 324 std::set<BasicBlock*> Visited; 325 Visited.insert(L->getHeader()); // Branches to header are ok. 326 BasicBlock *ExitBB = 0; 327 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 328 return ExitBB; 329 return 0; 330} 331 332/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 333/// trivial: that is, that the condition controls whether or not the loop does 334/// anything at all. If this is a trivial condition, unswitching produces no 335/// code duplications (equivalently, it produces a simpler loop and a new empty 336/// loop, which gets deleted). 337/// 338/// If this is a trivial condition, return true, otherwise return false. When 339/// returning true, this sets Cond and Val to the condition that controls the 340/// trivial condition: when Cond dynamically equals Val, the loop is known to 341/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 342/// Cond == Val. 343/// 344bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 345 BasicBlock **LoopExit) { 346 BasicBlock *Header = currentLoop->getHeader(); 347 TerminatorInst *HeaderTerm = Header->getTerminator(); 348 LLVMContext &Context = Header->getContext(); 349 350 BasicBlock *LoopExitBB = 0; 351 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 352 // If the header block doesn't end with a conditional branch on Cond, we 353 // can't handle it. 354 if (!BI->isConditional() || BI->getCondition() != Cond) 355 return false; 356 357 // Check to see if a successor of the branch is guaranteed to go to the 358 // latch block or exit through a one exit block without having any 359 // side-effects. If so, determine the value of Cond that causes it to do 360 // this. 361 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 362 BI->getSuccessor(0)))) { 363 if (Val) *Val = ConstantInt::getTrue(Context); 364 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 365 BI->getSuccessor(1)))) { 366 if (Val) *Val = ConstantInt::getFalse(Context); 367 } 368 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 369 // If this isn't a switch on Cond, we can't handle it. 370 if (SI->getCondition() != Cond) return false; 371 372 // Check to see if a successor of the switch is guaranteed to go to the 373 // latch block or exit through a one exit block without having any 374 // side-effects. If so, determine the value of Cond that causes it to do 375 // this. Note that we can't trivially unswitch on the default case. 376 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 377 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 378 SI->getSuccessor(i)))) { 379 // Okay, we found a trivial case, remember the value that is trivial. 380 if (Val) *Val = SI->getCaseValue(i); 381 break; 382 } 383 } 384 385 // If we didn't find a single unique LoopExit block, or if the loop exit block 386 // contains phi nodes, this isn't trivial. 387 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 388 return false; // Can't handle this. 389 390 if (LoopExit) *LoopExit = LoopExitBB; 391 392 // We already know that nothing uses any scalar values defined inside of this 393 // loop. As such, we just have to check to see if this loop will execute any 394 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 395 // part of the loop that the code *would* execute. We already checked the 396 // tail, check the header now. 397 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 398 if (I->mayHaveSideEffects()) 399 return false; 400 return true; 401} 402 403/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 404/// we choose to unswitch current loop on the specified value. 405/// 406unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) { 407 // If the condition is trivial, always unswitch. There is no code growth for 408 // this case. 409 if (IsTrivialUnswitchCondition(LIC)) 410 return 0; 411 412 // FIXME: This is overly conservative because it does not take into 413 // consideration code simplification opportunities. 414 CodeMetrics Metrics; 415 for (Loop::block_iterator I = currentLoop->block_begin(), 416 E = currentLoop->block_end(); 417 I != E; ++I) 418 Metrics.analyzeBasicBlock(*I); 419 return Metrics.NumInsts; 420} 421 422/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 423/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 424/// unswitch the loop, reprocess the pieces, then return true. 425bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){ 426 427 initLoopData(); 428 Function *F = loopHeader->getParent(); 429 430 431 // Check to see if it would be profitable to unswitch current loop. 432 unsigned Cost = getLoopUnswitchCost(LoopCond); 433 434 // Do not do non-trivial unswitch while optimizing for size. 435 if (Cost && OptimizeForSize) 436 return false; 437 if (Cost && !F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize)) 438 return false; 439 440 if (Cost > Threshold) { 441 // FIXME: this should estimate growth by the amount of code shared by the 442 // resultant unswitched loops. 443 // 444 DEBUG(errs() << "NOT unswitching loop %" 445 << currentLoop->getHeader()->getName() << ", cost too high: " 446 << currentLoop->getBlocks().size() << "\n"); 447 return false; 448 } 449 450 Constant *CondVal; 451 BasicBlock *ExitBlock; 452 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 453 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 454 } else { 455 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 456 } 457 458 return true; 459} 460 461// RemapInstruction - Convert the instruction operands from referencing the 462// current values into those specified by ValueMap. 463// 464static inline void RemapInstruction(Instruction *I, 465 DenseMap<const Value *, Value*> &ValueMap) { 466 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 467 Value *Op = I->getOperand(op); 468 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 469 if (It != ValueMap.end()) Op = It->second; 470 I->setOperand(op, Op); 471 } 472} 473 474/// CloneLoop - Recursively clone the specified loop and all of its children, 475/// mapping the blocks with the specified map. 476static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 477 LoopInfo *LI, LPPassManager *LPM) { 478 Loop *New = new Loop(); 479 480 LPM->insertLoop(New, PL); 481 482 // Add all of the blocks in L to the new loop. 483 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 484 I != E; ++I) 485 if (LI->getLoopFor(*I) == L) 486 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 487 488 // Add all of the subloops to the new loop. 489 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 490 CloneLoop(*I, New, VM, LI, LPM); 491 492 return New; 493} 494 495/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 496/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 497/// code immediately before InsertPt. 498void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 499 BasicBlock *TrueDest, 500 BasicBlock *FalseDest, 501 Instruction *InsertPt) { 502 // Insert a conditional branch on LIC to the two preheaders. The original 503 // code is the true version and the new code is the false version. 504 Value *BranchVal = LIC; 505 if (!isa<ConstantInt>(Val) || 506 Val->getType() != Type::getInt1Ty(LIC->getContext())) 507 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); 508 else if (Val != ConstantInt::getTrue(Val->getContext())) 509 // We want to enter the new loop when the condition is true. 510 std::swap(TrueDest, FalseDest); 511 512 // Insert the new branch. 513 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 514 515 // If either edge is critical, split it. This helps preserve LoopSimplify 516 // form for enclosing loops. 517 SplitCriticalEdge(BI, 0, this); 518 SplitCriticalEdge(BI, 1, this); 519} 520 521/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 522/// condition in it (a cond branch from its header block to its latch block, 523/// where the path through the loop that doesn't execute its body has no 524/// side-effects), unswitch it. This doesn't involve any code duplication, just 525/// moving the conditional branch outside of the loop and updating loop info. 526void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 527 Constant *Val, 528 BasicBlock *ExitBlock) { 529 DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %" 530 << loopHeader->getName() << " [" << L->getBlocks().size() 531 << " blocks] in Function " << L->getHeader()->getParent()->getName() 532 << " on cond: " << *Val << " == " << *Cond << "\n"); 533 534 // First step, split the preheader, so that we know that there is a safe place 535 // to insert the conditional branch. We will change loopPreheader to have a 536 // conditional branch on Cond. 537 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 538 539 // Now that we have a place to insert the conditional branch, create a place 540 // to branch to: this is the exit block out of the loop that we should 541 // short-circuit to. 542 543 // Split this block now, so that the loop maintains its exit block, and so 544 // that the jump from the preheader can execute the contents of the exit block 545 // without actually branching to it (the exit block should be dominated by the 546 // loop header, not the preheader). 547 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 548 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 549 550 // Okay, now we have a position to branch from and a position to branch to, 551 // insert the new conditional branch. 552 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 553 loopPreheader->getTerminator()); 554 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 555 loopPreheader->getTerminator()->eraseFromParent(); 556 557 // We need to reprocess this loop, it could be unswitched again. 558 redoLoop = true; 559 560 // Now that we know that the loop is never entered when this condition is a 561 // particular value, rewrite the loop with this info. We know that this will 562 // at least eliminate the old branch. 563 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 564 ++NumTrivial; 565} 566 567/// SplitExitEdges - Split all of the edges from inside the loop to their exit 568/// blocks. Update the appropriate Phi nodes as we do so. 569void LoopUnswitch::SplitExitEdges(Loop *L, 570 const SmallVector<BasicBlock *, 8> &ExitBlocks) 571{ 572 573 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 574 BasicBlock *ExitBlock = ExitBlocks[i]; 575 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 576 pred_end(ExitBlock)); 577 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), 578 ".us-lcssa", this); 579 } 580} 581 582/// UnswitchNontrivialCondition - We determined that the loop is profitable 583/// to unswitch when LIC equal Val. Split it into loop versions and test the 584/// condition outside of either loop. Return the loops created as Out1/Out2. 585void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 586 Loop *L) { 587 Function *F = loopHeader->getParent(); 588 DEBUG(errs() << "loop-unswitch: Unswitching loop %" 589 << loopHeader->getName() << " [" << L->getBlocks().size() 590 << " blocks] in Function " << F->getName() 591 << " when '" << *Val << "' == " << *LIC << "\n"); 592 593 LoopBlocks.clear(); 594 NewBlocks.clear(); 595 596 // First step, split the preheader and exit blocks, and add these blocks to 597 // the LoopBlocks list. 598 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 599 LoopBlocks.push_back(NewPreheader); 600 601 // We want the loop to come after the preheader, but before the exit blocks. 602 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 603 604 SmallVector<BasicBlock*, 8> ExitBlocks; 605 L->getUniqueExitBlocks(ExitBlocks); 606 607 // Split all of the edges from inside the loop to their exit blocks. Update 608 // the appropriate Phi nodes as we do so. 609 SplitExitEdges(L, ExitBlocks); 610 611 // The exit blocks may have been changed due to edge splitting, recompute. 612 ExitBlocks.clear(); 613 L->getUniqueExitBlocks(ExitBlocks); 614 615 // Add exit blocks to the loop blocks. 616 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 617 618 // Next step, clone all of the basic blocks that make up the loop (including 619 // the loop preheader and exit blocks), keeping track of the mapping between 620 // the instructions and blocks. 621 NewBlocks.reserve(LoopBlocks.size()); 622 DenseMap<const Value*, Value*> ValueMap; 623 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 624 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 625 NewBlocks.push_back(New); 626 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 627 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L); 628 } 629 630 // Splice the newly inserted blocks into the function right before the 631 // original preheader. 632 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 633 NewBlocks[0], F->end()); 634 635 // Now we create the new Loop object for the versioned loop. 636 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 637 Loop *ParentLoop = L->getParentLoop(); 638 if (ParentLoop) { 639 // Make sure to add the cloned preheader and exit blocks to the parent loop 640 // as well. 641 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 642 } 643 644 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 645 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 646 // The new exit block should be in the same loop as the old one. 647 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 648 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 649 650 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 651 "Exit block should have been split to have one successor!"); 652 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 653 654 // If the successor of the exit block had PHI nodes, add an entry for 655 // NewExit. 656 PHINode *PN; 657 for (BasicBlock::iterator I = ExitSucc->begin(); 658 (PN = dyn_cast<PHINode>(I)); ++I) { 659 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 660 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 661 if (It != ValueMap.end()) V = It->second; 662 PN->addIncoming(V, NewExit); 663 } 664 } 665 666 // Rewrite the code to refer to itself. 667 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 668 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 669 E = NewBlocks[i]->end(); I != E; ++I) 670 RemapInstruction(I, ValueMap); 671 672 // Rewrite the original preheader to select between versions of the loop. 673 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 674 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 675 "Preheader splitting did not work correctly!"); 676 677 // Emit the new branch that selects between the two versions of this loop. 678 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 679 LPM->deleteSimpleAnalysisValue(OldBR, L); 680 OldBR->eraseFromParent(); 681 682 LoopProcessWorklist.push_back(NewLoop); 683 redoLoop = true; 684 685 // Now we rewrite the original code to know that the condition is true and the 686 // new code to know that the condition is false. 687 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 688 689 // It's possible that simplifying one loop could cause the other to be 690 // deleted. If so, don't simplify it. 691 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 692 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 693 694} 695 696/// RemoveFromWorklist - Remove all instances of I from the worklist vector 697/// specified. 698static void RemoveFromWorklist(Instruction *I, 699 std::vector<Instruction*> &Worklist) { 700 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 701 Worklist.end(), I); 702 while (WI != Worklist.end()) { 703 unsigned Offset = WI-Worklist.begin(); 704 Worklist.erase(WI); 705 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 706 } 707} 708 709/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 710/// program, replacing all uses with V and update the worklist. 711static void ReplaceUsesOfWith(Instruction *I, Value *V, 712 std::vector<Instruction*> &Worklist, 713 Loop *L, LPPassManager *LPM) { 714 DEBUG(errs() << "Replace with '" << *V << "': " << *I); 715 716 // Add uses to the worklist, which may be dead now. 717 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 718 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 719 Worklist.push_back(Use); 720 721 // Add users to the worklist which may be simplified now. 722 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 723 UI != E; ++UI) 724 Worklist.push_back(cast<Instruction>(*UI)); 725 LPM->deleteSimpleAnalysisValue(I, L); 726 RemoveFromWorklist(I, Worklist); 727 I->replaceAllUsesWith(V); 728 I->eraseFromParent(); 729 ++NumSimplify; 730} 731 732/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 733/// information, and remove any dead successors it has. 734/// 735void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 736 std::vector<Instruction*> &Worklist, 737 Loop *L) { 738 if (pred_begin(BB) != pred_end(BB)) { 739 // This block isn't dead, since an edge to BB was just removed, see if there 740 // are any easy simplifications we can do now. 741 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 742 // If it has one pred, fold phi nodes in BB. 743 while (isa<PHINode>(BB->begin())) 744 ReplaceUsesOfWith(BB->begin(), 745 cast<PHINode>(BB->begin())->getIncomingValue(0), 746 Worklist, L, LPM); 747 748 // If this is the header of a loop and the only pred is the latch, we now 749 // have an unreachable loop. 750 if (Loop *L = LI->getLoopFor(BB)) 751 if (loopHeader == BB && L->contains(Pred)) { 752 // Remove the branch from the latch to the header block, this makes 753 // the header dead, which will make the latch dead (because the header 754 // dominates the latch). 755 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 756 Pred->getTerminator()->eraseFromParent(); 757 new UnreachableInst(BB->getContext(), Pred); 758 759 // The loop is now broken, remove it from LI. 760 RemoveLoopFromHierarchy(L); 761 762 // Reprocess the header, which now IS dead. 763 RemoveBlockIfDead(BB, Worklist, L); 764 return; 765 } 766 767 // If pred ends in a uncond branch, add uncond branch to worklist so that 768 // the two blocks will get merged. 769 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 770 if (BI->isUnconditional()) 771 Worklist.push_back(BI); 772 } 773 return; 774 } 775 776 DEBUG(errs() << "Nuking dead block: " << *BB); 777 778 // Remove the instructions in the basic block from the worklist. 779 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 780 RemoveFromWorklist(I, Worklist); 781 782 // Anything that uses the instructions in this basic block should have their 783 // uses replaced with undefs. 784 // If I is not void type then replaceAllUsesWith undef. 785 // This allows ValueHandlers and custom metadata to adjust itself. 786 if (I->getType() != Type::getVoidTy(I->getContext())) 787 I->replaceAllUsesWith(UndefValue::get(I->getType())); 788 } 789 790 // If this is the edge to the header block for a loop, remove the loop and 791 // promote all subloops. 792 if (Loop *BBLoop = LI->getLoopFor(BB)) { 793 if (BBLoop->getLoopLatch() == BB) 794 RemoveLoopFromHierarchy(BBLoop); 795 } 796 797 // Remove the block from the loop info, which removes it from any loops it 798 // was in. 799 LI->removeBlock(BB); 800 801 802 // Remove phi node entries in successors for this block. 803 TerminatorInst *TI = BB->getTerminator(); 804 SmallVector<BasicBlock*, 4> Succs; 805 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 806 Succs.push_back(TI->getSuccessor(i)); 807 TI->getSuccessor(i)->removePredecessor(BB); 808 } 809 810 // Unique the successors, remove anything with multiple uses. 811 array_pod_sort(Succs.begin(), Succs.end()); 812 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 813 814 // Remove the basic block, including all of the instructions contained in it. 815 LPM->deleteSimpleAnalysisValue(BB, L); 816 BB->eraseFromParent(); 817 // Remove successor blocks here that are not dead, so that we know we only 818 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 819 // then getting removed before we revisit them, which is badness. 820 // 821 for (unsigned i = 0; i != Succs.size(); ++i) 822 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 823 // One exception is loop headers. If this block was the preheader for a 824 // loop, then we DO want to visit the loop so the loop gets deleted. 825 // We know that if the successor is a loop header, that this loop had to 826 // be the preheader: the case where this was the latch block was handled 827 // above and headers can only have two predecessors. 828 if (!LI->isLoopHeader(Succs[i])) { 829 Succs.erase(Succs.begin()+i); 830 --i; 831 } 832 } 833 834 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 835 RemoveBlockIfDead(Succs[i], Worklist, L); 836} 837 838/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 839/// become unwrapped, either because the backedge was deleted, or because the 840/// edge into the header was removed. If the edge into the header from the 841/// latch block was removed, the loop is unwrapped but subloops are still alive, 842/// so they just reparent loops. If the loops are actually dead, they will be 843/// removed later. 844void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 845 LPM->deleteLoopFromQueue(L); 846 RemoveLoopFromWorklist(L); 847} 848 849// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 850// the value specified by Val in the specified loop, or we know it does NOT have 851// that value. Rewrite any uses of LIC or of properties correlated to it. 852void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 853 Constant *Val, 854 bool IsEqual) { 855 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 856 857 // FIXME: Support correlated properties, like: 858 // for (...) 859 // if (li1 < li2) 860 // ... 861 // if (li1 > li2) 862 // ... 863 864 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 865 // selects, switches. 866 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 867 std::vector<Instruction*> Worklist; 868 LLVMContext &Context = Val->getContext(); 869 870 871 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 872 // in the loop with the appropriate one directly. 873 if (IsEqual || (isa<ConstantInt>(Val) && 874 Val->getType() == Type::getInt1Ty(Val->getContext()))) { 875 Value *Replacement; 876 if (IsEqual) 877 Replacement = Val; 878 else 879 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 880 !cast<ConstantInt>(Val)->getZExtValue()); 881 882 for (unsigned i = 0, e = Users.size(); i != e; ++i) 883 if (Instruction *U = cast<Instruction>(Users[i])) { 884 if (!L->contains(U->getParent())) 885 continue; 886 U->replaceUsesOfWith(LIC, Replacement); 887 Worklist.push_back(U); 888 } 889 } else { 890 // Otherwise, we don't know the precise value of LIC, but we do know that it 891 // is certainly NOT "Val". As such, simplify any uses in the loop that we 892 // can. This case occurs when we unswitch switch statements. 893 for (unsigned i = 0, e = Users.size(); i != e; ++i) 894 if (Instruction *U = cast<Instruction>(Users[i])) { 895 if (!L->contains(U->getParent())) 896 continue; 897 898 Worklist.push_back(U); 899 900 // If we know that LIC is not Val, use this info to simplify code. 901 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 902 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 903 if (SI->getCaseValue(i) == Val) { 904 // Found a dead case value. Don't remove PHI nodes in the 905 // successor if they become single-entry, those PHI nodes may 906 // be in the Users list. 907 908 // FIXME: This is a hack. We need to keep the successor around 909 // and hooked up so as to preserve the loop structure, because 910 // trying to update it is complicated. So instead we preserve the 911 // loop structure and put the block on a dead code path. 912 BasicBlock *Switch = SI->getParent(); 913 SplitEdge(Switch, SI->getSuccessor(i), this); 914 // Compute the successors instead of relying on the return value 915 // of SplitEdge, since it may have split the switch successor 916 // after PHI nodes. 917 BasicBlock *NewSISucc = SI->getSuccessor(i); 918 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 919 // Create an "unreachable" destination. 920 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 921 Switch->getParent(), 922 OldSISucc); 923 new UnreachableInst(Context, Abort); 924 // Force the new case destination to branch to the "unreachable" 925 // block while maintaining a (dead) CFG edge to the old block. 926 NewSISucc->getTerminator()->eraseFromParent(); 927 BranchInst::Create(Abort, OldSISucc, 928 ConstantInt::getTrue(Context), NewSISucc); 929 // Release the PHI operands for this edge. 930 for (BasicBlock::iterator II = NewSISucc->begin(); 931 PHINode *PN = dyn_cast<PHINode>(II); ++II) 932 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 933 UndefValue::get(PN->getType())); 934 // Tell the domtree about the new block. We don't fully update the 935 // domtree here -- instead we force it to do a full recomputation 936 // after the pass is complete -- but we do need to inform it of 937 // new blocks. 938 if (DT) 939 DT->addNewBlock(Abort, NewSISucc); 940 break; 941 } 942 } 943 } 944 945 // TODO: We could do other simplifications, for example, turning 946 // LIC == Val -> false. 947 } 948 } 949 950 SimplifyCode(Worklist, L); 951} 952 953/// SimplifyCode - Okay, now that we have simplified some instructions in the 954/// loop, walk over it and constant prop, dce, and fold control flow where 955/// possible. Note that this is effectively a very simple loop-structure-aware 956/// optimizer. During processing of this loop, L could very well be deleted, so 957/// it must not be used. 958/// 959/// FIXME: When the loop optimizer is more mature, separate this out to a new 960/// pass. 961/// 962void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 963 while (!Worklist.empty()) { 964 Instruction *I = Worklist.back(); 965 Worklist.pop_back(); 966 967 // Simple constant folding. 968 if (Constant *C = ConstantFoldInstruction(I, I->getContext())) { 969 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 970 continue; 971 } 972 973 // Simple DCE. 974 if (isInstructionTriviallyDead(I)) { 975 DEBUG(errs() << "Remove dead instruction '" << *I); 976 977 // Add uses to the worklist, which may be dead now. 978 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 979 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 980 Worklist.push_back(Use); 981 LPM->deleteSimpleAnalysisValue(I, L); 982 RemoveFromWorklist(I, Worklist); 983 I->eraseFromParent(); 984 ++NumSimplify; 985 continue; 986 } 987 988 // Special case hacks that appear commonly in unswitched code. 989 switch (I->getOpcode()) { 990 case Instruction::Select: 991 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 992 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, 993 LPM); 994 continue; 995 } 996 break; 997 case Instruction::And: 998 if (isa<ConstantInt>(I->getOperand(0)) && 999 // constant -> RHS 1000 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext())) 1001 cast<BinaryOperator>(I)->swapOperands(); 1002 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1003 if (CB->getType() == Type::getInt1Ty(I->getContext())) { 1004 if (CB->isOne()) // X & 1 -> X 1005 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1006 else // X & 0 -> 0 1007 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1008 continue; 1009 } 1010 break; 1011 case Instruction::Or: 1012 if (isa<ConstantInt>(I->getOperand(0)) && 1013 // constant -> RHS 1014 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext())) 1015 cast<BinaryOperator>(I)->swapOperands(); 1016 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1017 if (CB->getType() == Type::getInt1Ty(I->getContext())) { 1018 if (CB->isOne()) // X | 1 -> 1 1019 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1020 else // X | 0 -> X 1021 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1022 continue; 1023 } 1024 break; 1025 case Instruction::Br: { 1026 BranchInst *BI = cast<BranchInst>(I); 1027 if (BI->isUnconditional()) { 1028 // If BI's parent is the only pred of the successor, fold the two blocks 1029 // together. 1030 BasicBlock *Pred = BI->getParent(); 1031 BasicBlock *Succ = BI->getSuccessor(0); 1032 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1033 if (!SinglePred) continue; // Nothing to do. 1034 assert(SinglePred == Pred && "CFG broken"); 1035 1036 DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- " 1037 << Succ->getName() << "\n"); 1038 1039 // Resolve any single entry PHI nodes in Succ. 1040 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1041 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1042 1043 // Move all of the successor contents from Succ to Pred. 1044 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1045 Succ->end()); 1046 LPM->deleteSimpleAnalysisValue(BI, L); 1047 BI->eraseFromParent(); 1048 RemoveFromWorklist(BI, Worklist); 1049 1050 // If Succ has any successors with PHI nodes, update them to have 1051 // entries coming from Pred instead of Succ. 1052 Succ->replaceAllUsesWith(Pred); 1053 1054 // Remove Succ from the loop tree. 1055 LI->removeBlock(Succ); 1056 LPM->deleteSimpleAnalysisValue(Succ, L); 1057 Succ->eraseFromParent(); 1058 ++NumSimplify; 1059 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1060 // Conditional branch. Turn it into an unconditional branch, then 1061 // remove dead blocks. 1062 break; // FIXME: Enable. 1063 1064 DEBUG(errs() << "Folded branch: " << *BI); 1065 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1066 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1067 DeadSucc->removePredecessor(BI->getParent(), true); 1068 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1069 LPM->deleteSimpleAnalysisValue(BI, L); 1070 BI->eraseFromParent(); 1071 RemoveFromWorklist(BI, Worklist); 1072 ++NumSimplify; 1073 1074 RemoveBlockIfDead(DeadSucc, Worklist, L); 1075 } 1076 break; 1077 } 1078 } 1079 } 1080} 1081